Huawei Ascend 910C: China Plans 600,000 AI Chips in 2026

Huawei plans to manufacture approximately 600,000 units of its Ascend 910C AI chip in 2026, nearly doubling its current output. Including other Ascend models, up to 1.6 million dies may be distributed across China's AI sector this year.

What Is the Huawei Ascend 910C?

The Ascend 910C is Huawei's most capable domestically produced AI accelerator. It is built on SMIC's enhanced 7nm process node, compared to the 4nm TSMC node Nvidia uses for its B200 chips. Each 910C delivers roughly one-third the BF16 throughput of Nvidia's B200. Chinese AI developers compensate by running larger clusters — scaling horizontally instead of vertically.

The 910C's primary customers are Alibaba, Tencent, and DeepSeek, all of which have committed to using domestic Huawei hardware for at least some of their AI workloads, partly due to US export restrictions and partly due to government pressure to support domestic chip suppliers.

A next-generation chip, the Ascend 950PR, is planned for Q1 2026.

Why This Happened

US export controls restricted Nvidia's ability to sell H100, A100, and even cut-down export versions of these chips to China beginning in late 2022. Nvidia designed a lower-spec H800 and A800 specifically for the Chinese market that stayed within export limits. Those too were restricted in October 2023. The B200 and anything equivalent is now entirely off the table for Chinese buyers.

The result: China had no choice but to build. Huawei's semiconductor arm (HiSilicon) accelerated development of the Ascend line. SMIC, China's largest domestic foundry, invested in an enhanced 7nm process. Neither is as advanced as TSMC or as capable as Nvidia's chips, but they exist, they work, and they are now being produced at meaningful scale.

The Separate Stack Problem

China's AI hardware ecosystem is now effectively isolated from the global one. Nvidia's competitive advantage is not just its chips — it is the CUDA software ecosystem, libraries, and toolchain that tens of thousands of AI researchers and engineers have built workflows around for 15 years.

Huawei's alternative is CANN (Compute Architecture for Neural Networks), its own programming framework for the Ascend series. Chinese AI companies training on Ascend chips must use CANN or write adapters for PyTorch and other frameworks. This creates a fork in the AI development world: one side runs on CUDA/Nvidia, the other on CANN/Ascend.

For Chinese AI startups and researchers, this is a significant productivity drag. For Chinese AI companies building their own models — Alibaba's Qwen, ByteDance's Doubao, DeepSeek — it means maintaining two codebases or investing heavily in compatibility layers.

DeepSeek's Chip Problem

DeepSeek R2, which has been anticipated since early 2025, has reportedly been delayed partly because of issues training on Huawei Ascend chips. According to reporting from August 2025, DeepSeek returned to using Nvidia H800s (the older export-allowed chip) for critical training runs after encountering problems at scale with Ascend hardware.

This is revealing. Even a company that has explicitly built its reputation on hardware efficiency — DeepSeek R1 was notable for running competitively on fewer chips — has found the Ascend stack limiting for training frontier models. The 910C works for inference and smaller fine-tuning runs. Full pre-training of frontier models is where the gap with Nvidia is still substantial.

Ascend 910C vs Nvidia: What the Numbers Mean

The performance gap is real. But for inference — running a trained model in production — the gap matters less than for training. A cluster of 910C chips delivering responses to millions of users is not doing frontier training; it is doing matrix multiplications at scale, which the 910C handles adequately.

China's path is to train on Nvidia where possible (using chips already in China before restrictions tightened) and deploy on Ascend for inference at scale.

Key Takeaways

• 600,000 units — Huawei Ascend 910C production target for 2026, nearly double 2025 output
• 1/3 the B200 throughput — Ascend 910C BF16 performance relative to Nvidia's current flagship chip
• SMIC 7nm — the domestic process node China is using for AI chips, vs TSMC 4nm for Nvidia
• DeepSeek R2 delayed — reportedly due to issues training on Ascend hardware, returned to Nvidia H800s
• For developers: China's AI ecosystem is forking from the global one at the hardware layer. Models trained on Ascend/CANN may have different characteristics than CUDA-trained equivalents. If deploying AI in China, expect a separate hardware and software stack.
• What to watch: Huawei Ascend 950PR launch in Q1 2026 — whether it closes the performance gap with Nvidia's B200 or remains a one-third performance chip